Saturday, October 07, 2023

An Engineer does (AI) Art - the beginning

I've been playing with automated image generation for a long time, starting with various bits of CGI for CAD in the day job, unglamorous things like dithering for 4-, 8- and 16-bit images, more interesting things like using depth maps to make line drawings out of architectural renders; and also using it for amateur art - for example, at least one cover of the Alarums and Excursions APAzine c1990.

When I stopped doing that for a day-job, it was rather more arm's length - seeing sites like "In 20 Years" (whose predictions must surely soon be testable), some site that would take a picture and render it in the style of El Greco - image above creation date 1-AUG-2009; then more recently (3 years ago) another face-to-portrait website, which rendered randomly in a variety of artists' styles

So when Stable Diffusion became available on HuggingFace last autumn, my first thoughts were "new profile pics using photos less than a decade old", but also "if this is intelligent, and takes pure text input, maybe I can illustrate the stories I wrote back in college days".

The former was quickly successful, as the image I adopted for my GitHub account shows

The latter was less successful - trying to prompt a character who on first glance looks like they stepped out of a black and white photograph seemed to be beyond its comprehension. But by the end of a second day trying, I had a limited proof of concept

Noodling around on HuggingFace, I was aiming for just one thing, a plausible cover illustration; but given the lack of control, it was a case of just rolling dem bones, until one day in early January, I got something so close, yet still so far.

soon after which I said enough is enough, and installed the Automatic1111 WebUI, in CPU only mode on the workstation I'd meant as a dev box in retirement. It might take 10 minutes for a single 512x512 or equivalent image, but finally I could start actually working with the tools.

To be continued...

Upcoming episode preview : how I'd render those last two pictures today

No comments :